1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright 2012 Google, Inc.
5 * Foreground allocator code: allocate buckets from freelist, and allocate in
6 * sector granularity from writepoints.
8 * bch2_bucket_alloc() allocates a single bucket from a specific device.
10 * bch2_bucket_alloc_set() allocates one or more buckets from different devices
11 * in a given filesystem.
15 #include "alloc_background.h"
16 #include "alloc_foreground.h"
17 #include "backpointers.h"
18 #include "btree_iter.h"
19 #include "btree_update.h"
22 #include "buckets_waiting_for_journal.h"
25 #include "disk_groups.h"
31 #include "nocow_locking.h"
34 #include <linux/math64.h>
35 #include <linux/rculist.h>
36 #include <linux/rcupdate.h>
38 static void bch2_trans_mutex_lock_norelock(struct btree_trans *trans,
41 if (!mutex_trylock(lock)) {
42 bch2_trans_unlock(trans);
47 const char * const bch2_watermarks[] = {
55 * Open buckets represent a bucket that's currently being allocated from. They
58 * - They track buckets that have been partially allocated, allowing for
59 * sub-bucket sized allocations - they're used by the sector allocator below
61 * - They provide a reference to the buckets they own that mark and sweep GC
62 * can find, until the new allocation has a pointer to it inserted into the
65 * When allocating some space with the sector allocator, the allocation comes
66 * with a reference to an open bucket - the caller is required to put that
67 * reference _after_ doing the index update that makes its allocation reachable.
70 void bch2_reset_alloc_cursors(struct bch_fs *c)
76 for_each_member_device_rcu(ca, c, i, NULL)
81 static void bch2_open_bucket_hash_add(struct bch_fs *c, struct open_bucket *ob)
83 open_bucket_idx_t idx = ob - c->open_buckets;
84 open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
90 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
92 open_bucket_idx_t idx = ob - c->open_buckets;
93 open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
95 while (*slot != idx) {
97 slot = &c->open_buckets[*slot].hash;
104 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
106 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
109 ec_stripe_new_put(c, ob->ec, STRIPE_REF_io);
113 percpu_down_read(&c->mark_lock);
114 spin_lock(&ob->lock);
119 spin_unlock(&ob->lock);
120 percpu_up_read(&c->mark_lock);
122 spin_lock(&c->freelist_lock);
123 bch2_open_bucket_hash_remove(c, ob);
125 ob->freelist = c->open_buckets_freelist;
126 c->open_buckets_freelist = ob - c->open_buckets;
128 c->open_buckets_nr_free++;
129 ca->nr_open_buckets--;
130 spin_unlock(&c->freelist_lock);
132 closure_wake_up(&c->open_buckets_wait);
135 void bch2_open_bucket_write_error(struct bch_fs *c,
136 struct open_buckets *obs,
139 struct open_bucket *ob;
142 open_bucket_for_each(c, obs, ob, i)
143 if (ob->dev == dev && ob->ec)
144 bch2_ec_bucket_cancel(c, ob);
147 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
149 struct open_bucket *ob;
151 BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
153 ob = c->open_buckets + c->open_buckets_freelist;
154 c->open_buckets_freelist = ob->freelist;
155 atomic_set(&ob->pin, 1);
158 c->open_buckets_nr_free--;
162 static void open_bucket_free_unused(struct bch_fs *c, struct open_bucket *ob)
164 BUG_ON(c->open_buckets_partial_nr >=
165 ARRAY_SIZE(c->open_buckets_partial));
167 spin_lock(&c->freelist_lock);
168 ob->on_partial_list = true;
169 c->open_buckets_partial[c->open_buckets_partial_nr++] =
170 ob - c->open_buckets;
171 spin_unlock(&c->freelist_lock);
173 closure_wake_up(&c->open_buckets_wait);
174 closure_wake_up(&c->freelist_wait);
177 /* _only_ for allocating the journal on a new device: */
178 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
180 while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
181 u64 b = ca->new_fs_bucket_idx++;
183 if (!is_superblock_bucket(ca, b) &&
184 (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
191 static inline unsigned open_buckets_reserved(enum bch_watermark watermark)
194 case BCH_WATERMARK_reclaim:
196 case BCH_WATERMARK_btree:
197 case BCH_WATERMARK_btree_copygc:
198 return OPEN_BUCKETS_COUNT / 4;
199 case BCH_WATERMARK_copygc:
200 return OPEN_BUCKETS_COUNT / 3;
202 return OPEN_BUCKETS_COUNT / 2;
206 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
208 enum bch_watermark watermark,
209 const struct bch_alloc_v4 *a,
210 struct bucket_alloc_state *s,
213 struct open_bucket *ob;
215 if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
220 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
225 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
226 c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
227 s->skipped_need_journal_commit++;
231 if (bch2_bucket_nocow_is_locked(&c->nocow_locks, POS(ca->dev_idx, bucket))) {
236 spin_lock(&c->freelist_lock);
238 if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(watermark))) {
240 closure_wait(&c->open_buckets_wait, cl);
242 track_event_change(&c->times[BCH_TIME_blocked_allocate_open_bucket],
243 &c->blocked_allocate_open_bucket, true);
244 spin_unlock(&c->freelist_lock);
245 return ERR_PTR(-BCH_ERR_open_buckets_empty);
248 /* Recheck under lock: */
249 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
250 spin_unlock(&c->freelist_lock);
255 ob = bch2_open_bucket_alloc(c);
257 spin_lock(&ob->lock);
260 ob->sectors_free = ca->mi.bucket_size;
261 ob->dev = ca->dev_idx;
264 spin_unlock(&ob->lock);
266 ca->nr_open_buckets++;
267 bch2_open_bucket_hash_add(c, ob);
269 track_event_change(&c->times[BCH_TIME_blocked_allocate_open_bucket],
270 &c->blocked_allocate_open_bucket, false);
272 track_event_change(&c->times[BCH_TIME_blocked_allocate],
273 &c->blocked_allocate, false);
275 spin_unlock(&c->freelist_lock);
279 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
280 enum bch_watermark watermark, u64 free_entry,
281 struct bucket_alloc_state *s,
282 struct bkey_s_c freespace_k,
285 struct bch_fs *c = trans->c;
286 struct btree_iter iter = { NULL };
288 struct open_bucket *ob;
289 struct bch_alloc_v4 a_convert;
290 const struct bch_alloc_v4 *a;
291 u64 b = free_entry & ~(~0ULL << 56);
292 unsigned genbits = free_entry >> 56;
293 struct printbuf buf = PRINTBUF;
296 if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
297 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
299 ca->mi.first_bucket, ca->mi.nbuckets);
300 bch2_bkey_val_to_text(&buf, c, freespace_k);
301 bch2_trans_inconsistent(trans, "%s", buf.buf);
306 k = bch2_bkey_get_iter(trans, &iter,
307 BTREE_ID_alloc, POS(ca->dev_idx, b),
315 a = bch2_alloc_to_v4(k, &a_convert);
317 if (a->data_type != BCH_DATA_free) {
318 if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
323 prt_printf(&buf, "non free bucket in freespace btree\n"
325 bch2_bkey_val_to_text(&buf, c, freespace_k);
326 prt_printf(&buf, "\n ");
327 bch2_bkey_val_to_text(&buf, c, k);
328 bch2_trans_inconsistent(trans, "%s", buf.buf);
333 if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
334 c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
335 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
337 genbits, alloc_freespace_genbits(*a) >> 56);
338 bch2_bkey_val_to_text(&buf, c, freespace_k);
339 prt_printf(&buf, "\n ");
340 bch2_bkey_val_to_text(&buf, c, k);
341 bch2_trans_inconsistent(trans, "%s", buf.buf);
346 if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_extents_to_backpointers) {
347 struct bch_backpointer bp;
348 struct bpos bp_pos = POS_MIN;
350 ret = bch2_get_next_backpointer(trans, POS(ca->dev_idx, b), -1,
352 BTREE_ITER_NOPRESERVE);
358 if (!bkey_eq(bp_pos, POS_MAX)) {
360 * Bucket may have data in it - we don't call
361 * bc2h_trans_inconnsistent() because fsck hasn't
369 ob = __try_alloc_bucket(c, ca, b, watermark, a, s, cl);
371 set_btree_iter_dontneed(&iter);
374 set_btree_iter_dontneed(&iter);
375 bch2_trans_iter_exit(trans, &iter);
381 * This path is for before the freespace btree is initialized:
383 * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
384 * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
386 static noinline struct open_bucket *
387 bch2_bucket_alloc_early(struct btree_trans *trans,
389 enum bch_watermark watermark,
390 struct bucket_alloc_state *s,
393 struct btree_iter iter, citer;
394 struct bkey_s_c k, ck;
395 struct open_bucket *ob = NULL;
396 u64 first_bucket = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
397 u64 alloc_start = max(first_bucket, READ_ONCE(ca->alloc_cursor));
398 u64 alloc_cursor = alloc_start;
402 * Scan with an uncached iterator to avoid polluting the key cache. An
403 * uncached iter will return a cached key if one exists, but if not
404 * there is no other underlying protection for the associated key cache
405 * slot. To avoid racing bucket allocations, look up the cached key slot
406 * of any likely allocation candidate before attempting to proceed with
407 * the allocation. This provides proper exclusion on the associated
411 for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, alloc_cursor),
412 BTREE_ITER_SLOTS, k, ret) {
413 struct bch_alloc_v4 a_convert;
414 const struct bch_alloc_v4 *a;
416 if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
419 if (ca->new_fs_bucket_idx &&
420 is_superblock_bucket(ca, k.k->p.offset))
423 a = bch2_alloc_to_v4(k, &a_convert);
424 if (a->data_type != BCH_DATA_free)
427 /* now check the cached key to serialize concurrent allocs of the bucket */
428 ck = bch2_bkey_get_iter(trans, &citer, BTREE_ID_alloc, k.k->p, BTREE_ITER_CACHED);
433 a = bch2_alloc_to_v4(ck, &a_convert);
434 if (a->data_type != BCH_DATA_free)
439 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, watermark, a, s, cl);
441 set_btree_iter_dontneed(&citer);
442 bch2_trans_iter_exit(trans, &citer);
446 bch2_trans_iter_exit(trans, &iter);
448 alloc_cursor = iter.pos.offset;
449 ca->alloc_cursor = alloc_cursor;
454 if (!ob && alloc_start > first_bucket) {
455 alloc_cursor = alloc_start = first_bucket;
462 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
464 enum bch_watermark watermark,
465 struct bucket_alloc_state *s,
468 struct btree_iter iter;
470 struct open_bucket *ob = NULL;
471 u64 alloc_start = max_t(u64, ca->mi.first_bucket, READ_ONCE(ca->alloc_cursor));
472 u64 alloc_cursor = alloc_start;
475 BUG_ON(ca->new_fs_bucket_idx);
477 for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
478 POS(ca->dev_idx, alloc_cursor), 0, k, ret) {
479 if (k.k->p.inode != ca->dev_idx)
482 for (alloc_cursor = max(alloc_cursor, bkey_start_offset(k.k));
483 alloc_cursor < k.k->p.offset;
485 ret = btree_trans_too_many_iters(trans);
493 ob = try_alloc_bucket(trans, ca, watermark,
494 alloc_cursor, s, k, cl);
496 set_btree_iter_dontneed(&iter);
504 bch2_trans_iter_exit(trans, &iter);
506 ca->alloc_cursor = alloc_cursor;
511 if (!ob && alloc_start > ca->mi.first_bucket) {
512 alloc_cursor = alloc_start = ca->mi.first_bucket;
520 * bch2_bucket_alloc_trans - allocate a single bucket from a specific device
521 * @trans: transaction object
522 * @ca: device to allocate from
523 * @watermark: how important is this allocation?
524 * @cl: if not NULL, closure to be used to wait if buckets not available
525 * @usage: for secondarily also returning the current device usage
527 * Returns: an open_bucket on success, or an ERR_PTR() on failure.
529 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
531 enum bch_watermark watermark,
533 struct bch_dev_usage *usage)
535 struct bch_fs *c = trans->c;
536 struct open_bucket *ob = NULL;
537 bool freespace = READ_ONCE(ca->mi.freespace_initialized);
539 struct bucket_alloc_state s = { 0 };
540 bool waiting = false;
542 bch2_dev_usage_read_fast(ca, usage);
543 avail = dev_buckets_free(ca, *usage, watermark);
545 if (usage->d[BCH_DATA_need_discard].buckets > avail)
548 if (usage->d[BCH_DATA_need_gc_gens].buckets > avail)
551 if (should_invalidate_buckets(ca, *usage))
552 bch2_do_invalidates(c);
555 if (cl && !waiting) {
556 closure_wait(&c->freelist_wait, cl);
561 track_event_change(&c->times[BCH_TIME_blocked_allocate],
562 &c->blocked_allocate, true);
564 ob = ERR_PTR(-BCH_ERR_freelist_empty);
569 closure_wake_up(&c->freelist_wait);
571 ob = likely(freespace)
572 ? bch2_bucket_alloc_freelist(trans, ca, watermark, &s, cl)
573 : bch2_bucket_alloc_early(trans, ca, watermark, &s, cl);
575 if (s.skipped_need_journal_commit * 2 > avail)
576 bch2_journal_flush_async(&c->journal, NULL);
578 if (!ob && freespace && c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
584 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
587 trace_and_count(c, bucket_alloc, ca,
588 bch2_watermarks[watermark],
590 usage->d[BCH_DATA_free].buckets,
592 bch2_copygc_wait_amount(c),
593 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
597 else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
598 trace_and_count(c, bucket_alloc_fail, ca,
599 bch2_watermarks[watermark],
601 usage->d[BCH_DATA_free].buckets,
603 bch2_copygc_wait_amount(c),
604 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
607 bch2_err_str(PTR_ERR(ob)));
612 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
613 enum bch_watermark watermark,
616 struct bch_dev_usage usage;
617 struct open_bucket *ob;
619 bch2_trans_do(c, NULL, NULL, 0,
620 PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(trans, ca, watermark,
625 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
626 unsigned l, unsigned r)
628 return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
629 (stripe->next_alloc[l] < stripe->next_alloc[r]));
632 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
634 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
635 struct dev_stripe_state *stripe,
636 struct bch_devs_mask *devs)
638 struct dev_alloc_list ret = { .nr = 0 };
641 for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
642 ret.devs[ret.nr++] = i;
644 bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
648 static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
649 struct dev_stripe_state *stripe,
650 struct bch_dev_usage *usage)
652 u64 *v = stripe->next_alloc + ca->dev_idx;
653 u64 free_space = dev_buckets_available(ca, BCH_WATERMARK_normal);
654 u64 free_space_inv = free_space
655 ? div64_u64(1ULL << 48, free_space)
659 if (*v + free_space_inv >= *v)
660 *v += free_space_inv;
664 for (v = stripe->next_alloc;
665 v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
666 *v = *v < scale ? 0 : *v - scale;
669 void bch2_dev_stripe_increment(struct bch_dev *ca,
670 struct dev_stripe_state *stripe)
672 struct bch_dev_usage usage;
674 bch2_dev_usage_read_fast(ca, &usage);
675 bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
678 static int add_new_bucket(struct bch_fs *c,
679 struct open_buckets *ptrs,
680 struct bch_devs_mask *devs_may_alloc,
681 unsigned nr_replicas,
682 unsigned *nr_effective,
685 struct open_bucket *ob)
687 unsigned durability =
688 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
690 BUG_ON(*nr_effective >= nr_replicas);
692 __clear_bit(ob->dev, devs_may_alloc->d);
693 *nr_effective += durability;
694 *have_cache |= !durability;
696 ob_push(c, ptrs, ob);
698 if (*nr_effective >= nr_replicas)
705 int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
706 struct open_buckets *ptrs,
707 struct dev_stripe_state *stripe,
708 struct bch_devs_mask *devs_may_alloc,
709 unsigned nr_replicas,
710 unsigned *nr_effective,
713 enum bch_data_type data_type,
714 enum bch_watermark watermark,
717 struct bch_fs *c = trans->c;
718 struct dev_alloc_list devs_sorted =
719 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
722 int ret = -BCH_ERR_insufficient_devices;
725 BUG_ON(*nr_effective >= nr_replicas);
727 for (i = 0; i < devs_sorted.nr; i++) {
728 struct bch_dev_usage usage;
729 struct open_bucket *ob;
731 dev = devs_sorted.devs[i];
734 ca = rcu_dereference(c->devs[dev]);
736 percpu_ref_get(&ca->ref);
742 if (!ca->mi.durability && *have_cache) {
743 percpu_ref_put(&ca->ref);
747 ob = bch2_bucket_alloc_trans(trans, ca, watermark, cl, &usage);
749 bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
750 percpu_ref_put(&ca->ref);
754 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
759 ob->data_type = data_type;
761 if (add_new_bucket(c, ptrs, devs_may_alloc,
762 nr_replicas, nr_effective,
763 have_cache, flags, ob)) {
772 /* Allocate from stripes: */
775 * if we can't allocate a new stripe because there are already too many
776 * partially filled stripes, force allocating from an existing stripe even when
777 * it's to a device we don't want:
780 static int bucket_alloc_from_stripe(struct btree_trans *trans,
781 struct open_buckets *ptrs,
782 struct write_point *wp,
783 struct bch_devs_mask *devs_may_alloc,
785 unsigned nr_replicas,
786 unsigned *nr_effective,
788 enum bch_watermark watermark,
792 struct bch_fs *c = trans->c;
793 struct dev_alloc_list devs_sorted;
794 struct ec_stripe_head *h;
795 struct open_bucket *ob;
802 if (ec_open_bucket(c, ptrs))
805 h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, watermark, cl);
811 devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
813 for (i = 0; i < devs_sorted.nr; i++)
814 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
815 if (!h->s->blocks[ec_idx])
818 ob = c->open_buckets + h->s->blocks[ec_idx];
819 if (ob->dev == devs_sorted.devs[i] &&
820 !test_and_set_bit(ec_idx, h->s->blocks_allocated))
827 ec_stripe_new_get(h->s, STRIPE_REF_io);
829 ret = add_new_bucket(c, ptrs, devs_may_alloc,
830 nr_replicas, nr_effective,
831 have_cache, flags, ob);
833 bch2_ec_stripe_head_put(c, h);
837 /* Sector allocator */
839 static bool want_bucket(struct bch_fs *c,
840 struct write_point *wp,
841 struct bch_devs_mask *devs_may_alloc,
842 bool *have_cache, bool ec,
843 struct open_bucket *ob)
845 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
847 if (!test_bit(ob->dev, devs_may_alloc->d))
850 if (ob->data_type != wp->data_type)
853 if (!ca->mi.durability &&
854 (wp->data_type == BCH_DATA_btree || ec || *have_cache))
857 if (ec != (ob->ec != NULL))
863 static int bucket_alloc_set_writepoint(struct bch_fs *c,
864 struct open_buckets *ptrs,
865 struct write_point *wp,
866 struct bch_devs_mask *devs_may_alloc,
867 unsigned nr_replicas,
868 unsigned *nr_effective,
870 bool ec, unsigned flags)
872 struct open_buckets ptrs_skip = { .nr = 0 };
873 struct open_bucket *ob;
877 open_bucket_for_each(c, &wp->ptrs, ob, i) {
878 if (!ret && want_bucket(c, wp, devs_may_alloc,
880 ret = add_new_bucket(c, ptrs, devs_may_alloc,
881 nr_replicas, nr_effective,
882 have_cache, flags, ob);
884 ob_push(c, &ptrs_skip, ob);
886 wp->ptrs = ptrs_skip;
891 static int bucket_alloc_set_partial(struct bch_fs *c,
892 struct open_buckets *ptrs,
893 struct write_point *wp,
894 struct bch_devs_mask *devs_may_alloc,
895 unsigned nr_replicas,
896 unsigned *nr_effective,
897 bool *have_cache, bool ec,
898 enum bch_watermark watermark,
903 if (!c->open_buckets_partial_nr)
906 spin_lock(&c->freelist_lock);
908 if (!c->open_buckets_partial_nr)
911 for (i = c->open_buckets_partial_nr - 1; i >= 0; --i) {
912 struct open_bucket *ob = c->open_buckets + c->open_buckets_partial[i];
914 if (want_bucket(c, wp, devs_may_alloc, have_cache, ec, ob)) {
915 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
916 struct bch_dev_usage usage;
919 bch2_dev_usage_read_fast(ca, &usage);
920 avail = dev_buckets_free(ca, usage, watermark);
924 array_remove_item(c->open_buckets_partial,
925 c->open_buckets_partial_nr,
927 ob->on_partial_list = false;
929 ret = add_new_bucket(c, ptrs, devs_may_alloc,
930 nr_replicas, nr_effective,
931 have_cache, flags, ob);
937 spin_unlock(&c->freelist_lock);
941 static int __open_bucket_add_buckets(struct btree_trans *trans,
942 struct open_buckets *ptrs,
943 struct write_point *wp,
944 struct bch_devs_list *devs_have,
947 unsigned nr_replicas,
948 unsigned *nr_effective,
950 enum bch_watermark watermark,
954 struct bch_fs *c = trans->c;
955 struct bch_devs_mask devs;
956 struct open_bucket *ob;
957 struct closure *cl = NULL;
961 devs = target_rw_devs(c, wp->data_type, target);
963 /* Don't allocate from devices we already have pointers to: */
964 for (i = 0; i < devs_have->nr; i++)
965 __clear_bit(devs_have->devs[i], devs.d);
967 open_bucket_for_each(c, ptrs, ob, i)
968 __clear_bit(ob->dev, devs.d);
970 if (erasure_code && ec_open_bucket(c, ptrs))
973 ret = bucket_alloc_set_writepoint(c, ptrs, wp, &devs,
974 nr_replicas, nr_effective,
975 have_cache, erasure_code, flags);
979 ret = bucket_alloc_set_partial(c, ptrs, wp, &devs,
980 nr_replicas, nr_effective,
981 have_cache, erasure_code, watermark, flags);
986 ret = bucket_alloc_from_stripe(trans, ptrs, wp, &devs,
988 nr_replicas, nr_effective,
990 watermark, flags, _cl);
994 * Try nonblocking first, so that if one device is full we'll try from
997 ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
998 nr_replicas, nr_effective, have_cache,
999 flags, wp->data_type, watermark, cl);
1001 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
1002 !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
1005 goto retry_blocking;
1012 static int open_bucket_add_buckets(struct btree_trans *trans,
1013 struct open_buckets *ptrs,
1014 struct write_point *wp,
1015 struct bch_devs_list *devs_have,
1017 unsigned erasure_code,
1018 unsigned nr_replicas,
1019 unsigned *nr_effective,
1021 enum bch_watermark watermark,
1028 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1029 devs_have, target, erasure_code,
1030 nr_replicas, nr_effective, have_cache,
1031 watermark, flags, cl);
1032 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1033 bch2_err_matches(ret, BCH_ERR_operation_blocked) ||
1034 bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
1035 bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1037 if (*nr_effective >= nr_replicas)
1041 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1042 devs_have, target, false,
1043 nr_replicas, nr_effective, have_cache,
1044 watermark, flags, cl);
1045 return ret < 0 ? ret : 0;
1049 * should_drop_bucket - check if this is open_bucket should go away
1050 * @ob: open_bucket to predicate on
1051 * @c: filesystem handle
1052 * @ca: if set, we're killing buckets for a particular device
1053 * @ec: if true, we're shutting down erasure coding and killing all ec
1055 * otherwise, return true
1056 * Returns: true if we should kill this open_bucket
1058 * We're killing open_buckets because we're shutting down a device, erasure
1059 * coding, or the entire filesystem - check if this open_bucket matches:
1061 static bool should_drop_bucket(struct open_bucket *ob, struct bch_fs *c,
1062 struct bch_dev *ca, bool ec)
1065 return ob->ec != NULL;
1067 bool drop = ob->dev == ca->dev_idx;
1068 struct open_bucket *ob2;
1071 if (!drop && ob->ec) {
1074 mutex_lock(&ob->ec->lock);
1075 nr_blocks = bkey_i_to_stripe(&ob->ec->new_stripe.key)->v.nr_blocks;
1077 for (i = 0; i < nr_blocks; i++) {
1078 if (!ob->ec->blocks[i])
1081 ob2 = c->open_buckets + ob->ec->blocks[i];
1082 drop |= ob2->dev == ca->dev_idx;
1084 mutex_unlock(&ob->ec->lock);
1093 static void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
1094 bool ec, struct write_point *wp)
1096 struct open_buckets ptrs = { .nr = 0 };
1097 struct open_bucket *ob;
1100 mutex_lock(&wp->lock);
1101 open_bucket_for_each(c, &wp->ptrs, ob, i)
1102 if (should_drop_bucket(ob, c, ca, ec))
1103 bch2_open_bucket_put(c, ob);
1105 ob_push(c, &ptrs, ob);
1107 mutex_unlock(&wp->lock);
1110 void bch2_open_buckets_stop(struct bch_fs *c, struct bch_dev *ca,
1115 /* Next, close write points that point to this device... */
1116 for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1117 bch2_writepoint_stop(c, ca, ec, &c->write_points[i]);
1119 bch2_writepoint_stop(c, ca, ec, &c->copygc_write_point);
1120 bch2_writepoint_stop(c, ca, ec, &c->rebalance_write_point);
1121 bch2_writepoint_stop(c, ca, ec, &c->btree_write_point);
1123 mutex_lock(&c->btree_reserve_cache_lock);
1124 while (c->btree_reserve_cache_nr) {
1125 struct btree_alloc *a =
1126 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1128 bch2_open_buckets_put(c, &a->ob);
1130 mutex_unlock(&c->btree_reserve_cache_lock);
1132 spin_lock(&c->freelist_lock);
1134 while (i < c->open_buckets_partial_nr) {
1135 struct open_bucket *ob =
1136 c->open_buckets + c->open_buckets_partial[i];
1138 if (should_drop_bucket(ob, c, ca, ec)) {
1139 --c->open_buckets_partial_nr;
1140 swap(c->open_buckets_partial[i],
1141 c->open_buckets_partial[c->open_buckets_partial_nr]);
1142 ob->on_partial_list = false;
1143 spin_unlock(&c->freelist_lock);
1144 bch2_open_bucket_put(c, ob);
1145 spin_lock(&c->freelist_lock);
1150 spin_unlock(&c->freelist_lock);
1152 bch2_ec_stop_dev(c, ca);
1155 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
1156 unsigned long write_point)
1159 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
1161 return &c->write_points_hash[hash];
1164 static struct write_point *__writepoint_find(struct hlist_head *head,
1165 unsigned long write_point)
1167 struct write_point *wp;
1170 hlist_for_each_entry_rcu(wp, head, node)
1171 if (wp->write_point == write_point)
1179 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1181 u64 stranded = c->write_points_nr * c->bucket_size_max;
1182 u64 free = bch2_fs_usage_read_short(c).free;
1184 return stranded * factor > free;
1187 static bool try_increase_writepoints(struct bch_fs *c)
1189 struct write_point *wp;
1191 if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1192 too_many_writepoints(c, 32))
1195 wp = c->write_points + c->write_points_nr++;
1196 hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1200 static bool try_decrease_writepoints(struct btree_trans *trans, unsigned old_nr)
1202 struct bch_fs *c = trans->c;
1203 struct write_point *wp;
1204 struct open_bucket *ob;
1207 mutex_lock(&c->write_points_hash_lock);
1208 if (c->write_points_nr < old_nr) {
1209 mutex_unlock(&c->write_points_hash_lock);
1213 if (c->write_points_nr == 1 ||
1214 !too_many_writepoints(c, 8)) {
1215 mutex_unlock(&c->write_points_hash_lock);
1219 wp = c->write_points + --c->write_points_nr;
1221 hlist_del_rcu(&wp->node);
1222 mutex_unlock(&c->write_points_hash_lock);
1224 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1225 open_bucket_for_each(c, &wp->ptrs, ob, i)
1226 open_bucket_free_unused(c, ob);
1228 mutex_unlock(&wp->lock);
1232 static struct write_point *writepoint_find(struct btree_trans *trans,
1233 unsigned long write_point)
1235 struct bch_fs *c = trans->c;
1236 struct write_point *wp, *oldest;
1237 struct hlist_head *head;
1239 if (!(write_point & 1UL)) {
1240 wp = (struct write_point *) write_point;
1241 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1245 head = writepoint_hash(c, write_point);
1247 wp = __writepoint_find(head, write_point);
1250 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1251 if (wp->write_point == write_point)
1253 mutex_unlock(&wp->lock);
1256 restart_find_oldest:
1258 for (wp = c->write_points;
1259 wp < c->write_points + c->write_points_nr; wp++)
1260 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1263 bch2_trans_mutex_lock_norelock(trans, &oldest->lock);
1264 bch2_trans_mutex_lock_norelock(trans, &c->write_points_hash_lock);
1265 if (oldest >= c->write_points + c->write_points_nr ||
1266 try_increase_writepoints(c)) {
1267 mutex_unlock(&c->write_points_hash_lock);
1268 mutex_unlock(&oldest->lock);
1269 goto restart_find_oldest;
1272 wp = __writepoint_find(head, write_point);
1273 if (wp && wp != oldest) {
1274 mutex_unlock(&c->write_points_hash_lock);
1275 mutex_unlock(&oldest->lock);
1280 hlist_del_rcu(&wp->node);
1281 wp->write_point = write_point;
1282 hlist_add_head_rcu(&wp->node, head);
1283 mutex_unlock(&c->write_points_hash_lock);
1285 wp->last_used = local_clock();
1289 static noinline void
1290 deallocate_extra_replicas(struct bch_fs *c,
1291 struct open_buckets *ptrs,
1292 struct open_buckets *ptrs_no_use,
1293 unsigned extra_replicas)
1295 struct open_buckets ptrs2 = { 0 };
1296 struct open_bucket *ob;
1299 open_bucket_for_each(c, ptrs, ob, i) {
1300 unsigned d = bch_dev_bkey_exists(c, ob->dev)->mi.durability;
1302 if (d && d <= extra_replicas) {
1303 extra_replicas -= d;
1304 ob_push(c, ptrs_no_use, ob);
1306 ob_push(c, &ptrs2, ob);
1314 * Get us an open_bucket we can allocate from, return with it locked:
1316 int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1318 unsigned erasure_code,
1319 struct write_point_specifier write_point,
1320 struct bch_devs_list *devs_have,
1321 unsigned nr_replicas,
1322 unsigned nr_replicas_required,
1323 enum bch_watermark watermark,
1326 struct write_point **wp_ret)
1328 struct bch_fs *c = trans->c;
1329 struct write_point *wp;
1330 struct open_bucket *ob;
1331 struct open_buckets ptrs;
1332 unsigned nr_effective, write_points_nr;
1337 if (!IS_ENABLED(CONFIG_BCACHEFS_ERASURE_CODING))
1338 erasure_code = false;
1340 BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
1342 BUG_ON(!nr_replicas || !nr_replicas_required);
1346 write_points_nr = c->write_points_nr;
1349 *wp_ret = wp = writepoint_find(trans, write_point.v);
1351 /* metadata may not allocate on cache devices: */
1352 if (wp->data_type != BCH_DATA_user)
1355 if (target && !(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1356 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1357 target, erasure_code,
1358 nr_replicas, &nr_effective,
1359 &have_cache, watermark,
1362 bch2_err_matches(ret, BCH_ERR_transaction_restart))
1365 /* Don't retry from all devices if we're out of open buckets: */
1366 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1367 goto allocate_blocking;
1370 * Only try to allocate cache (durability = 0 devices) from the
1375 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1377 nr_replicas, &nr_effective,
1378 &have_cache, watermark,
1382 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1383 target, erasure_code,
1384 nr_replicas, &nr_effective,
1385 &have_cache, watermark,
1389 BUG_ON(!ret && nr_effective < nr_replicas);
1391 if (erasure_code && !ec_open_bucket(c, &ptrs))
1392 pr_debug("failed to get ec bucket: ret %u", ret);
1394 if (ret == -BCH_ERR_insufficient_devices &&
1395 nr_effective >= nr_replicas_required)
1401 if (nr_effective > nr_replicas)
1402 deallocate_extra_replicas(c, &ptrs, &wp->ptrs, nr_effective - nr_replicas);
1404 /* Free buckets we didn't use: */
1405 open_bucket_for_each(c, &wp->ptrs, ob, i)
1406 open_bucket_free_unused(c, ob);
1410 wp->sectors_free = UINT_MAX;
1412 open_bucket_for_each(c, &wp->ptrs, ob, i)
1413 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1415 BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1419 open_bucket_for_each(c, &wp->ptrs, ob, i)
1420 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1421 ob_push(c, &ptrs, ob);
1423 open_bucket_free_unused(c, ob);
1426 mutex_unlock(&wp->lock);
1428 if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1429 try_decrease_writepoints(trans, write_points_nr))
1432 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1433 bch2_err_matches(ret, BCH_ERR_freelist_empty))
1435 ? -BCH_ERR_bucket_alloc_blocked
1436 : -BCH_ERR_ENOSPC_bucket_alloc;
1441 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1443 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1445 return (struct bch_extent_ptr) {
1446 .type = 1 << BCH_EXTENT_ENTRY_ptr,
1449 .offset = bucket_to_sector(ca, ob->bucket) +
1450 ca->mi.bucket_size -
1455 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1456 struct bkey_i *k, unsigned sectors,
1459 bch2_alloc_sectors_append_ptrs_inlined(c, wp, k, sectors, cached);
1463 * Append pointers to the space we just allocated to @k, and mark @sectors space
1464 * as allocated out of @ob
1466 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1468 bch2_alloc_sectors_done_inlined(c, wp);
1471 static inline void writepoint_init(struct write_point *wp,
1472 enum bch_data_type type)
1474 mutex_init(&wp->lock);
1475 wp->data_type = type;
1477 INIT_WORK(&wp->index_update_work, bch2_write_point_do_index_updates);
1478 INIT_LIST_HEAD(&wp->writes);
1479 spin_lock_init(&wp->writes_lock);
1482 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1484 struct open_bucket *ob;
1485 struct write_point *wp;
1487 mutex_init(&c->write_points_hash_lock);
1488 c->write_points_nr = ARRAY_SIZE(c->write_points);
1490 /* open bucket 0 is a sentinal NULL: */
1491 spin_lock_init(&c->open_buckets[0].lock);
1493 for (ob = c->open_buckets + 1;
1494 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1495 spin_lock_init(&ob->lock);
1496 c->open_buckets_nr_free++;
1498 ob->freelist = c->open_buckets_freelist;
1499 c->open_buckets_freelist = ob - c->open_buckets;
1502 writepoint_init(&c->btree_write_point, BCH_DATA_btree);
1503 writepoint_init(&c->rebalance_write_point, BCH_DATA_user);
1504 writepoint_init(&c->copygc_write_point, BCH_DATA_user);
1506 for (wp = c->write_points;
1507 wp < c->write_points + c->write_points_nr; wp++) {
1508 writepoint_init(wp, BCH_DATA_user);
1510 wp->last_used = local_clock();
1511 wp->write_point = (unsigned long) wp;
1512 hlist_add_head_rcu(&wp->node,
1513 writepoint_hash(c, wp->write_point));
1517 static void bch2_open_bucket_to_text(struct printbuf *out, struct bch_fs *c, struct open_bucket *ob)
1519 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1520 unsigned data_type = ob->data_type;
1521 barrier(); /* READ_ONCE() doesn't work on bitfields */
1523 prt_printf(out, "%zu ref %u %s %u:%llu gen %u allocated %u/%u",
1524 ob - c->open_buckets,
1525 atomic_read(&ob->pin),
1526 data_type < BCH_DATA_NR ? bch2_data_types[data_type] : "invalid data type",
1527 ob->dev, ob->bucket, ob->gen,
1528 ca->mi.bucket_size - ob->sectors_free, ca->mi.bucket_size);
1530 prt_printf(out, " ec idx %llu", ob->ec->idx);
1531 if (ob->on_partial_list)
1532 prt_str(out, " partial");
1536 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1538 struct open_bucket *ob;
1542 for (ob = c->open_buckets;
1543 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1545 spin_lock(&ob->lock);
1546 if (ob->valid && !ob->on_partial_list)
1547 bch2_open_bucket_to_text(out, c, ob);
1548 spin_unlock(&ob->lock);
1554 void bch2_open_buckets_partial_to_text(struct printbuf *out, struct bch_fs *c)
1559 spin_lock(&c->freelist_lock);
1561 for (i = 0; i < c->open_buckets_partial_nr; i++)
1562 bch2_open_bucket_to_text(out, c,
1563 c->open_buckets + c->open_buckets_partial[i]);
1565 spin_unlock(&c->freelist_lock);
1569 static const char * const bch2_write_point_states[] = {
1571 WRITE_POINT_STATES()
1576 static void bch2_write_point_to_text(struct printbuf *out, struct bch_fs *c,
1577 struct write_point *wp)
1579 struct open_bucket *ob;
1582 prt_printf(out, "%lu: ", wp->write_point);
1583 prt_human_readable_u64(out, wp->sectors_allocated);
1585 prt_printf(out, " last wrote: ");
1586 bch2_pr_time_units(out, sched_clock() - wp->last_used);
1588 for (i = 0; i < WRITE_POINT_STATE_NR; i++) {
1589 prt_printf(out, " %s: ", bch2_write_point_states[i]);
1590 bch2_pr_time_units(out, wp->time[i]);
1595 printbuf_indent_add(out, 2);
1596 open_bucket_for_each(c, &wp->ptrs, ob, i)
1597 bch2_open_bucket_to_text(out, c, ob);
1598 printbuf_indent_sub(out, 2);
1601 void bch2_write_points_to_text(struct printbuf *out, struct bch_fs *c)
1603 struct write_point *wp;
1605 prt_str(out, "Foreground write points\n");
1606 for (wp = c->write_points;
1607 wp < c->write_points + ARRAY_SIZE(c->write_points);
1609 bch2_write_point_to_text(out, c, wp);
1611 prt_str(out, "Copygc write point\n");
1612 bch2_write_point_to_text(out, c, &c->copygc_write_point);
1614 prt_str(out, "Rebalance write point\n");
1615 bch2_write_point_to_text(out, c, &c->rebalance_write_point);
1617 prt_str(out, "Btree write point\n");
1618 bch2_write_point_to_text(out, c, &c->btree_write_point);